def _get_candidate(self, parameters, results, lower_is_better,
min_candidates=10):
"""
Samples candidates parameters from the top 33% of population. If less than min_candidates
then use a random sample
Returns
dict: parameter dictionary.
"""
if results.shape[0] > 0:
population = results.loc[results['Status'] != 'INTERMEDIATE',
:] # select only completed trials
else:
population = None
if population is None or population.shape[0] < min_candidates:
trial_param_values = {}
for parameter_object in parameters:
trial_param_values[
parameter_object.name] = parameter_object.sample()
return trial_param_values
population = population.sort_values(by='Objective',
ascending=lower_is_better)
idx = sherpa_rng.randint(low=0, high=population.shape[
0] // 3) # pick randomly among top 33%
trial_all_values = population.iloc[
idx].to_dict() # extract the trial values on results table
trial_param_values = {param.name: trial_all_values[param.name] for param
in parameters} # Select only parameter values
return trial_param_values
def _truncation_selection(self, parameters, results, lower_is_better):
"""
Continues the top 80% of the generation, resamples the rest from the
top 20% and perturbs.
Returns
dict: parameter dictionary.
"""
# Select correct generation and sort generation members
completed = results.loc[results['Status'] == 'COMPLETED', :]
if (self.count - 1) % self.population_size / self.population_size < 0.8:
# Go through top 80% of generation
generation_df = completed.loc[(completed.generation
== self.generation - 1), :] \
.sort_values(by='Objective',
ascending=lower_is_better)
d = generation_df.iloc[(self.count - 1) % self.population_size].to_dict()
else:
# For the rest, sample from top 20% of the last generation
target_generations = [self.generation - 1]
generation_df = completed.loc[(completed.generation.isin(
target_generations)), :] \
.sort_values(by='Objective',
ascending=lower_is_better)
idx = sherpa_rng.randint(low=0, high=self.population_size*len(target_generations)//5)
d = generation_df.iloc[idx].to_dict()
d = self._perturb(candidate=d, parameters=parameters)
trial = {param.name: d[param.name] for param in parameters}
for key in ['load_from', 'save_to', 'lineage']:
trial[key] = d[key]
return trial